On then Use of Windows with the Discrete

FREDRIC

for Harmonic Analysis Fourier Transform

MEXBER, IEEE

J. HARRIS,

HERE IS MUCH signal processing devoted to detection and estimation. Detection is the task of detetmitdng if a specific signal set is pteaettt in an obs&tion, whflc estimation is the task of obtaining the va.iues of the parameters derriblng the signal. Often the s@tal is complicated or is corrupted by interfeting signals or noise To facilitate the detection and estimation of signal sets. the obsenation is decomposed by a basis set which spans the signal space [ 1) For many problems of engineering interest, the class of aigttlls being sought are periodic which leads quite natuallv to a decomposition by a basis consistittg of simple petiodic fun=tions, the sines and cosines. The classic Fourier tran.,fot,,, h the mechanism by which we M able to perform this decomposttmn. BY necessity, every observed signal we pmmust be of finite extent. The extent may be adjustable and Axtable. but it must be fire. Proces%ng a fiite-duration observation ~POSCS mteresting and interacting considentior,s on the hamomc analysic rhese consldentions include detectability of tones in the Presence of nearby strong tones, rcoohability of similarstrength nearby tones, tesolvability of Gxifting tona, and biases in estimating the parameten of my of the alonmenhoned signals. For practicality, the data we pare N unifomdy spaced samples of the obsetvcd signal. For convenience. N is highJy composite, and we will zwtme N is evett. The harmottic estm~afes we obtain UtmugJt the discrae Fowie~ tmnsfotm (DFT) arc N mifcwmly spaced samples of the asaciated periodic spectra. This approach in elegant and attnctive when the proce~ scheme is cast as a spectral decomposition in an N-dimensional orthogonal vector space 121. Unfottunately, in mmY practical situations, to obtain meaningful results this elegance must be compmmised. One such

t=O,l;..,N-

l.N.N+l.

(1)
We cherve that by defining a buds set ovet an ordered index t, we am defining the rpcctmm over a line (called the frequen=Y Uu) from Which we dnW the concepts of bandindth md of frwuencia close to and fat from a given frequency (rhich is related to l-e%htion). For nmpled s@als. the basix set spatming the interval of NT scxxmdr h identid with the scqucnces obtained by uniform satnpka of the wmsponding wntinuous spanning Set up IO the index N/2. k=O,,.-.-.N/2 n=O.,;--.N/ (2) We note hen that the trigonometric functions are unique in that uniformly spxed samples (ovet an integer number of periods) fotm ortlt~ottal sequences. Arbitrary onhogond functions, zimiluly sampied, do not form orthogonal sequences. We also note that aa interval of length NT seconds 1s not the same as the inteml coveted by N samples separated by intervals of T seconds. This is easily understood when we by U.S. copytight

U.S. Government

work not protected

reali.ze that the interval oveq which the samples are t*cn is closed on the left and is open on the right (i.e., I-)). Fig. 1 demonstntes this by sampling a function which k even about its midpoint and of duration hT sccondr. Since the DFT essentiaUy considers squences to be periodic, scqcnee. of we can consider the mirnng end point to be the beming the next period of the periodic extension of this sequence. In is decomposed into its even and odd parts. with the odd part fact, under the periodic extensiott, the next sample (at 16 sin supplying the imaginary sine component in the ftite Fii. 1.) is indM&gGhable from the sample at zero seconds. transform. This apparent lack of symmetry due to the missing (but III. SPECTRAL LEAKAGE implied) end point is P +ource of umfusion in sampled window The selection of a ftite-time interval of NT seconds and of de&n. Thk can be traced to the early work related to conthe orthogonal trigonometric basis (continuous or sampled) vergence factors for the partial sums of the Fourier series. The over this interval leads to an interesting pectitiy of the partial sums (or the Gnite Fourier transform) always include spectral expansion. From the continuum of possible frean odd number of paints and exhibit even symmetry about quencies. only those which coincide with the basis will protect the origin. Hence much of the litekrc and many software onto a single basis vector; all other frequencies will exhibtt libr.uia incnrponte windows de-signed with true even symnon zero projections on the entire basis set. This 1s often metry rather than the implied symmetry with the missing end referred to as spectral leakage and is the result of processing point! of sampled data that We must remember for DFT prwe&ng ftite-duration records. Although the amount of leakage 1s influenced by the sampling period, leakage Is not caused by even aymmctry means that the projection upon the sampled the sampling. sine sequences is identically zero; it doa not mean a matching left and right data point about the midpoint. To distinguish An intuitive approach to leakage is the undentandmg that thi3 symmetry from cawentiond ~emten we wiU refer to it simals with frequencies other than those of the basis set are as DFT-even (i.e., a cmwentio~I even sequence with the rightnot periodic in the observation window. The petiodic exfension of a signal not commensurate with its natural period end point removed). Another example of DFT-even symexhibits discontiwities at the boundaCes of the observatmn. metry is presented in Fii. 2 aa samples of a periodicaIly extended triangle wave. The discontinuities are responsiblr for spectral contributions (61 leakage) over the entire basis xl The fo~rn) ni ihis du If WC evaluate a DFT-even sequence via a finite Fourier transform (by treating the +N/Z point as a zero-value point), contmuity are demonstrated in Fig. 1. the resultant continuous periodic function exhibits a non zero Windows are weighting functions applied to data to reduce imaginary component. The DFT of the same wquenre is a set the spectral leakage associated with fmite observation interof samples of the tiite Fourier transform, yet thex samples vals From one viewpoint, the window is applied to data exhibit an imaginary component equal to ICTO. Why the dis(a.5 a multiplicative weighting) to reduce the order of the di.parity? We mlut remember that the mixing end point under contlnurty at the boundary of the periodic extension. This is the DFT symmetry contributes an imaginary smusoidal accomplished by matching as many orders of derivative (of component of period Zn/(N/Z) to the finite transform the we&red data) as possible at the boundary. The easiest (corresponding to the odd component at sequence position way to achieve this matchmg is by setting the value of these N/2). The sampling positions of the DFT are at the multiples denvarwes 10 zero or near to zero. Thus windowed data are of 277/N. which, of course. correspond to the zeros of the smoothly brought to zero at the boundaries so that the im&m, sinuoidal component. An example of this forperiodic extension of the data is continuous in many orders tu~tou sampling is show in Fig. 3. Notice the sequencei( of derivative.

( i ! I 1 I , ! I I

HARRIS:

SE

OF WlNDOWS

FOR

HARMONIC

ANALYSIS

where

+=Zk. Nl

nndk=O,l;.~,N-

From another viewpoint, the window is mukiplioativtdy applied to the basis set so that a signal of arbitrary frequency will exhibit a ugnificanf projection only oo those basis rectors having a frequency close to the signal frequency. Of cause both viewpoints lead to identical results. We can w insight into window design by occasionaUy switching between these Vi.Wp0inti. IV. WINDOWS ANND FIGURFS OF MERIT Windows are wed in harmonic pnalysk to reduce the undesirable effects related to spectnl le. Windows implct on many attributes of a harmonic prooeso r; these indude detectability, resolution, dynamic range. confidence, and ease of implementation. We would like to identify the major paruneten that will allow perfcrmance comparisons between dilfennt windows. We can best identify the= pawneten by examining the effects on harmonic analysis of a window. An essentially bandlimited signal f(f) with Fourier transform F(w) can be described by the uniformly sampled data set /(nT). This data set defmes the periodically extended rpectrum FT(m) by its Fourier series expansion as idenas +F(w) = I__ f(r) exp (-jut) df (3s)

We recognize (4s) ~1 the finite Fourin tnnafonn, , summt ticm addrcratd for the corwenience of its even symmetry. E.qution (4b) is the finite Fourier transform with the xi&tend Point dckted. and (4~) is the DFT ymplin# of (4b). Of c&lfor sctui proaaing, PC desire (for count@ purpaa in a4withms) that the index start .t zero. We accomplkb this by sbifk the start& point of the data N/2 poC tions. ,zhawiw (4~) to(dd). Equation (4d) is the forward DFT lie N/2 shift will affect only the phue an&s of the tnns form. a for th? amvw.dence of symmetry we will address tbc window 11 being centered St the origin. We alao identify tbi? convenkaa u rmjor source of window misappliuhon. The shift of N/2 pc4nia and its dtmt phase shift is often overlooked or h impmpCay ham&d in the definition of the wiodor when usd with the DFT. This is putintkrly so when the window@ is performed as. s~echal convolution. See the dismion on the Fbmin# window under the cm= (1) windows. The question nor Poled is. to what sxtent is the fink@ sumrmtion of (4b) a manin@ approxinutim of the in6nite summatim of (3b)? In facts we addrcr the question for c more encnl case of m arbitnry window applird to the time function (or &ried) a.5 plxxntcd in F,(u) where w(nT)=O. and w(n7-J = w(-nn, L.et us *ov e umine the effects of the window on o,u Jpcctnl estimate. Equlioa (5) shhoan that the transfoml F,(w) ti the transform of . product. As indicated Ur the fouowing equation, ttds is eqtivaknt to the con*olution of the two comxponding tnnsforms (see Appendix): F,(w) =
F(x) W(w x)dx/2n

= i7 a=--

ww9mn

exp (-iwnn

(5)

N even

F(w)

+m = 1 =-_ +-IT = IF(

f(nn

exp (-iwIn

(3b)

f(f)

I -?f,T = 0, = F(U),

FT(w)

exp (+;a)

dwl2n

(3.9

and where

Iwllfl2n/n IWI c + IZnlTl.

FT(W)

For (real-world) machine process&, the data most be of finite extent, and the rummrtion of (3b) can only by performed as a fmte approximation as indicated a;i F,(w)= +N 2 i /(nneexP(-iwnn =-N/2
(NP-1
F,(w)

/- _m

(6)

F(w)

W(w)

Neven

(41)

Fb(W)=

z ((n77ecxp(-jwn77 n = -N/2 (NI2l-L

= x /(nT) exp (-;w~n77,

Neven

(4b)

Equation (6) is the key to the effects of procoed& rtite. extent data. The equation an he interpreted in two cquiralent waya, ahic,, wi be more c,,i,y vinulired with the aid of an cxampk. The example we choose U the sampled rectanzle wiodowa; ~(7) = 1.0. We know W(w) is the h&let kernel (4 I presec.ted as UT
1 sin u. N TUT

F,(w,)

Never,

(4~1 W(w) =exp

(

=-N,l Fdwc) .v 1 = z n=o

J(nT)

exp (-;w*nT).

N even

(4d)

1 sin 1 TWT I

(7)

except for the linear phase shift term (which will change due to the N,2 point shift for realizability), a single period of the transform has the form indicated in Fig. 5. The observation mmm,~ (6) is that the value of F,(w) at a particular W, say w - weI is the sum of ali of the spectral contributions at each w wcigbted by the window centered at we and measured at w (see Fig. 6). A. Equivalent Noise Bandwidth From Fii. 6, we observe @at the amplitude of the harmonic estimate at P given frzqoency is biased by the accumulated broad-band noise included in the bandwidth of the window. In this ywe, the wiodoa behaves a9 a Nter, gathering conmbutiona for its estimate over its bandwidth. For the harmonic detection problem, we de&e to minimize this accumulated noise &oat, md we accomplish this with small-bandwidth window. A cawenient measure of this baodwidth is the equivalent noise bandwidth (ENBW) of the window. This is the width of a ncta,&e ftiter with the same pcok power sain that would accumtitc the same noise power (se-e Fii. 7). The xcumolated noise power of the window is dcfmed as +-IT Noise Power = No I -,T IW(w)l dw/?n

(8)
Pamevas Nter is matched to one of the complex sinusoidal sequencer of the basis set [ 31. From this perspective, ,ve can examme the PC (sometimes called the coherent gain) of the filter. and we can examic the PL due to the window having reduced the data to zero vplues near the boundaries. Let the i,,put sampled sequence be defmed by (12): /(nT) (lob) = A CXP (+;wrnT) +&IT) IIZJ

where No Is the noise power pet unit bandwidth. theorem allows (8) to be computed by Noixe Power = F The peak power gain of the window frequency power gain. and is defined 2 w(nT). n

occurs at w = 0. the zero by (lOa) ~

Peak Signal Gain * W(0) - x w(nT) n PcakPowerGaioW(O)=

[n 1
c w(nT) 2

Tbur the ENBW (normalized by N,/T, the noise power per bin) is given in the following equation and is tabulated for the windows of this report in Table I

where q(nT) is a white-noise sequence with variance 0:. Then the sign.4 component of the windowed spectrum (the matche<: filler output) i3 presented in Rw*) /,ipul = 2 w(nJ7 A exp (+iw,n7) n
=A x w(nT).

exp (-w.+nT) (13)

z w(nT)
(I I)

A coocept closely tied to ENBW is procesiog gaiz IPG) and processing loss (PL) of a windowed transform We can think of the DFT as a bank of matched fdten, where each

We see that the noiseles measurement (the expected value of the nosy measurement) is proportional to the input amplitude .A The prownionabty factor is the sum of the window terms. which U in fact the dc sigrm, gain of the window. For a rectangle wndow tbh factor is N, the number of term m the window. For any other window, the gain is reduced due to the window smoothly going to zero near the boundaries. This

__-

-q-T -K-p7
I I

56

reduction in proportionality factor is important as It reprenents a known bias on spectral amplitudes. Coherent power gain. the square of coherent gain. is occasion& the parameter listed in the literature. Coherent gain (the summation of (13)) normalized by its maximum value N is listed in Table I. ,-be incoherent component of the windowed transfon is ,$ven by w(nT)q(nT) exp (-ianT) (14a) 7%) jnc.iac =x n and the incoherent power (the meansquarr value of this co,panent where E { } is the expectation operator) is given by
E{IF(wr)/.,,Il = 2 x w(nT)w(mT)EIq(nT)q*(mr)) n m

exp (-jw*nT) = 0; x w(nr). n

exp (+jwrmT)
rY-: e-1 intervals.

(I%)

Fig. 9.

Relationship

between

indices

on Overhppcd

Notice the incoherent power gain is the sum of the squares of the window tertes, and the coherent power gain is the square of the sum of the window terms. Finally, PS, which is dcfqxd as the ratio of output s&naltcmoise ratio to input signal-tcmoise ratio, is given.by pG~S,,N, Wi -A [? w(nr)]p
A/O:,

&

w(nT)

A,, imp,,rtant question related to overlapped processing is what is the degree of comlation of the random components in successie transforms? This cnm1ation, as a function of fractional overlap r. is deftned for a relatively flat noise spectrum OYCT the window bandwidth by (17). Fig. 9 identifies how the indices of (I 7) relate to the overlap of the intervals. The correlation coefficient

cwcnn 2 .[;: 1 nw(nr)

Notice PG is the reciprocal of the normalized ENBW. Thus large ENBW suggesta a reduced processing gain. This is reasonable, tice an increased noise bandwidth permits additional noise to contribute to a spectral estimate. c. Over17pConelation When the fast Fourier transform (FFT) is used to process long-time sequences a partition length N is fust selected to estabbsb the required spectral re-sc~lution of the enolysis. Spectral resolution of the FFT is defied in (16) where Af is the resoltion, f, ls the sample frequency selected to latify the Nyqut criterion. end 0 is the cafficient rcflecfing the bandwidth increuc due to the perticulu window selected. Note that [f,/N] is the minimum te.wlution of the FFf which we denote as the FFf bin width. The unfficicnt ,4 is usually selected to be the ENBW in bina u listed in Table I

is computed and tabulated in Table 1. for each of the windows listed for SO-and 75.percent overlap. Often in a spectral analysis. the squared magnitude of successive ttansfotms are averaged to reduce the va,iance of the measu~ements (S] We know of cotse that when we average K identically disttibutcd independent measurements. the vatiawe of the average is related to the individual vatiansr of the measurements by kg. - __t 181 diea K Now we can ask what is the reduction in the variance when we avenge measuemenfs which are correlated as they are for ovcrfappcd transforms? Welch 151 has supplied an answer to this question which we present here, for the special case of SOand 75.percent overlap dn,. 1. %ea =Kl + ?c(O.S)] iz; lc(OS)], 50 percent overlap =;ll+2c $ (0.75)+ 2c(o.s)+ lc(O.75)+ Zc(O.2S)l

Af=fl ; 0

(16)

If the window and the FFT are applied to nonovcr,apping partitions of the sequence. as shhown in Fig. 8. a slgruricant part of the reties is ignored due to the windows exhibiting small values near the boundaks. Far instance. if the transform IS being used to detect rho,+duration tone-like signals. the non overlapped analysrs could rmss the event if it occurred neu the boudancs. To avoid this loss of data. the transfons ax usually applied to the overlapped panition sequences as shown in Fig. 8. The overlap LS almost always 50 or 75 percent. Thx overlap processing of course mcreases the work load to cover the total sequence length, but the rewards warrant the extra effort.

2~(0.5)+3~(0.25j1. 75 percent overlap. (19)

The negative terms in (19) are the edge effects of the average and can be ignoted if the number of temu K is larger than ten. For good windows. ~(0.25) Is small campared ,a 1.0.

and can also be omitted from (19) with negligible error. For this reason, ~(0.25) was not listed in Table I. Note, that for good windows (see last paragraph ot Section IV-F), transforms taken with SO-percent overlap are essentially independent. D. Scalloping Loss

An important consideration related to minimum detectable signal is called scallopiag lass or picket-fence effect. We have considered the windowed DFT as a bank of matched ffiten and have examined the procexing gain and the reduction of this gain ascribable to thb window for tones matched to the basis vectols. The b&s vectors are tones with frequencies equal to multiples of f,/N (with f, being the sample frequency). These frequencies are sample points from the spectrum. and are nommlly referred to as DFT wtpt points or as DFT bins. We now address the question, what is the additional loss in procasing gain for a tone of frequency midway between two bin frequencies (that is, at frequencies

be used. Additional comments on poor windwvx wil! h, found in Section N-C. We can wncludr from tL< comb: ij~ loss fi of Table I and from Fig. 1: that for the deterria?:, of single tones in broad-band noise. nearly any window C(AIIC~ than the rectangle) is as gcal LS any other. The difference between the various windows is less than 1.0 dB and for good windows is Icar than 0.7 dB. The detection of t&s in the presence of other tones is, however, quite another problem. Here the window dccs have P marked affect, as will be demonstrated shcdy. F. Specml Leakwe Rrvidted Returning to (6) and to Fig. 6, we observe the spectral measlnment is affected not only by the broadhand coispectrum, but also by the narrow-band spatrun: rhic!~ ::ti!: within the bandwidth of the window. In fact, a given spia~i companent say .t w - wg will contribute output (or will bc observed) at mother frequency. say .t w = w. according to the gain of the window centered .t wo and measured at w.,. This is the effect normrlly referred to u spectral leakage and Is demonstrated in Fii. 10 with the transform of a finite duntion tone of frequency wg This leakage cases P bias in the amplitude and the position of P harmonic estimate. Even for the case of a single real harmonic line (not at a DFT sample point), the leakage from the kernel on the negative-frequency axis biases the kernel on the positive-frequency brie. This bias is most severe and 10s: bothersome for the detection of small signals in the presence of nearby huge signals. To reduce the effects of this bias, the window should exhibit low-amplitude sidelobes far from the central main Lobe, and the transition to the low sidelobes should be very rapid. One indicator of how well a window suppresses leakage h the peak sidelobe level (relative to the main lobe): another is the asymptotic rate of falloff of thesp sidelobes. TIew indicators are listed in Table 1.

(k+ ll2)f,lW
i S

Returning to (13), mie the processing defined in

with w* replaced by w(~+~,~), v/e detergain for this hall-bin frequency shift as ,,,,nT),

FkJ( I,I)) IsigN! = A x w(n73 exp (-iw( n

We ah de!ine the scalloping loss as the ratio of coherent gain for a tone located half a bin from a DFT sample point to the coherent gain for a tone located at a DFT sample point, as indicated in

IT w(nn=w(-jin)( lw(+F)!
scalloping Las = 1 w(n?-l n = WJ) (2Ob) Scalloping loss represents the maximun reduction in PC due to signal frequency. This loss has been computed for the windows of this report and has been included in Table I.

We now make an mtererting observation. We define worst case PL as the sun, of maximum sca,,oping loss of a window and of PL due to that wmdow (both in decibel). This number is the reduction of output s&m&to-noise ratio as a result of windowing and of wont case frequency location. This of course Is related to the minimum detectable tone in broadband noise. It IS interesting to note that the wont case loss is ahays between 3.0 and 4.3 dB. Windows with wont case PL exceeding 3.8 dB are very poor windows and should not

Fig. I1 suggests another criterion with which w%. should: 1, concerned in the window selection process. Since the win<ii>\v imposes an effective bandwidth on the spectral line, we would be interested in the minimum separation between two equalstrength lines such that for arbitrary spectral locations their respective main lobes can be resolved. The classic criterion for this resolution is the width of the window at the half-power points (the 3.0.dB bandwidth). This criterion reflects the fact that two equalstrength main lobes separated in frequency by less than their 3.0.dB bandwidths will exhibit a single spectral peak and will not be resolved as two distinct lines. The problem with this criterion is that it does not work for the coherent addition we find in the DFT. The DF? output points are the coherent addition of the spectral components weighted through the window at a given frequency.

If two kernels M contributing to the coherent summation, the sum at the crc6soer point bvxnidly half-way between them) must be smaller than the individual pea!~s if the two peaks arc to be resolved. Thus at the crossover points 01 fhc kernels, the gain from each kernel mwt be less than 0.5,01 the crcssovcr points must occur beyond the 6.0-dB points of the windoan. Table I lists the 6.C-dB bandwidths of the variOU window e xamined in this report. From the table, we see that the 6.0sdB bandwidth varies from 1.2 bins to 2.6 bins, where a bin is the fundamental frequency resolution w,/h. The 3.0-dB bandwidth does have utility as P pafommncc lniicator as shown in the next paragraph. Remember however. it is the 6.0-dB bandwidth which defmes the resolution of the win. dowed DFT. Fro,,, Table I, we see that the noise handwidth always exceeda the 3.0-dB bandwidth. The difference between the two, mfercnced to the 3.0-dB bandwidth, appears to be a sensitive indicator of overall window perfotmnnce. We have observed that for all the wad windows on the table, this indicator was found to be in the range of 4.0 to 5.5 percent. Tbae window8 for which this ratio is outside that range either have L wide main lobe or a bigb sidelobe structure and, hena. arc chancterized by hi& processing las.s or by poor two-tone detection capabilities. Those windows for which this mtio ia inside the 4.0 to 5.5perant range are found in i;.. the lower lcfturraa of the ~erformvla comparison chart 1; (Fe. 12). which is deaaibed next, .< While Table I dces list the common performance param;; cten of the windova e xamined in this report, the mass of : numben U not s~tenins. We do nti that the aidelobe :, kvd (to mducs hhs) and the wont case procrtsiag Ices (to xauindn dstsctabi&y) am probably the meat important paameten on the table. Fig. 12 shows the relative position of the tidorr u i hmtion of tbwc parameters. Windows rcaidiug in the lower left comer of the fwe are the good~~perforudng windgaa. They exhibit low-sidelobe levels and loss. We urge. the reader to read low aomt am proasdne ~seetioru VI and VU; Fii. 12 praents a lot of infonmtioo, but not the Ml Itmy. v. We wiII nc.w utdw know window% For L jutifiatim for ita ua _ AU the tidowa uiiJ c-c WrNwws ame reU-kaown (and some not ~weUeach window we will comment on the and identify Ita significant p-eten. be wkd aa even (about the ori@n) aeqtrsncu with m odd muhbex of Q&&S. To convert the windew to DFFeven. the t,&cnd point will be discarded and the seqcqucna dl be shIfted so that the left cod point coincidea with the origin. We uiUuss normalized coordinates with sample petiod r = I .o, So tb:ir w in pcnodic in 27r m, len<,i: will be identified 18 0. A DFT bin will be considered to extend between DFT sample points (multipleo of 2n,A, and have, width of 2nlN.
RrcrangIe IDirichkr) Window /6]

defined

as w(n)= 1.0, =O.I:~~,NI C:lhI

The spectral window

for the DFT window

W(S) = exp

The transform of this widow is seen to be the Dmchlcr kernel. which exhibits a DFT main-lobe width (between LCCTO crcssingx) of 2 bins and a fmt sidelobe level approum.wly I3 dB down from the main-lobe peak. The sidelobes fti off II 6.0 dB per octave, which is of course the expected rate lo, a function with P discontinuity. The parmeterr of the DFT window are listed in Table 1. question posed earlier: in what rense does tbr (224 aPProtiate the infinite sum of (22b)? +?v,z
F(B) = x =-N,I f(n) cxp (-inO) (::a)

N Eln Te NI [ 1 I:,<, ((2 1. sm [;a 1

-, -,g -___

sequence 1s @*en in

finrtc sun of

A.

The rectansle window is unity over the observation interval. md can be tholyht of as a gating sequence applied to the data so that they are of finite extent. Tbc window for a finire Fourier transform is defined as w(n)=l.O. and is shown in Fig. = L3. +1.0. 1, N 1

tF(B) = x n=-_

f(n)

exp (-inO).

(:?b)

(2la)

The same window

for a DFT is

We observe the tiite sum is the rectangle-windowed verxon of the infinite, mm. W recognize that~ the infinite rum is the Fouler smes ex~ansmn of some periodic function for which the I(n)? are the Foutier series coefficients. We also recognize -

59

that the finite sum is simply the partial sum of the series. From this viewpoint we can cast the question in terms of the convergence properties of the partial sums of Fouier series. From this work we know the partial sum is the least meansquare error approximation to the infmite sum We observe that mean square convergence IS a convement analytic concept, but it is not attractive for finite estimates ot ior numencd approximations. Meansquare estimates tend to orcdlate about their means. and do not exhibit umtorm convergence. (The approximation in a neighborhood of a point of contlnulty may get worse if more terms are added to the pamal rum.l We normally observe this behavior near pomtr of dlrcantmuity as the figin& we call Gibbs phenomenon. It is th,s orcibtory behavior we UC trying to control by the se of other wndows.

and the spectral window given in

corresponding

to the DFT sequeoce is

w@~=~exp

[-i(N;-

1) e]

W(n) = 1.o Jr?!

N/2 and ~1 shown defined as in Fig

N n=--,. 2
14.

-l.o,l.--.~

(23a) for a DFT 1s

The same wmdow

,
,i n

The transform of this window is seen to be the squared Dirichlet kernel. Its main-lobe width (between zero crossings) ix twice that of the nctan%er and the fu-st ndelobe level is approximately 26 dB down from the main-lobe peak. e&n. twxe that of the rectandcs. The sidelobes fall off at 12 dB per octwe, reflecting the discontinuity of the nndow residing in the first derivative (rather than in the funcuon Itulf, The triangle is the simplest window which exhihtts a nonneeatiw tranrftxw 3%~ property can be retired by ronvolvi,,.: a,, wmdaw (of half-extent) with itxlf. The resultant wi,dow! transform is the square of the o,i@nal windows tnnsfonn. A wndow sequence derived by self-conrolm a parent srindew contains approximately twice the number of samples as the parent window, hence curesponds to . tnwwxnetnc polynomiaJ (its Z-transform) of approximately twit-e the order. (Convol~ two recta,@es each of N/2 points wi result in a ttianglc of N + 1 points when the zero end points are counted.) The transform of the window will now exhibit twce ar many zeros as the parent transform (to account for the increased order of the associated trigonometric polynomial,. But how has the transform applied these extra zeros availabie from the increased order polynomial? The self-

of the cosine function. attractive under the DFT. transform is defied as

These propeties are particularly The wmdow for a finite Fourier

and for a DFT as

Notice the effect due to the change of the ori@x~ The most common values of a are the integer5 1 through 4. with 2 hew the most we known (as the Harming window). Tbis wmdou is identified for values of a equal to 1 and 2 in (263 iZk:,s (27a). and (27b), (the a for the finire transforms. the h for the DFT):

convolved window simply places repented zeros at each IOCation for which the parent transform had a zero. This, Of came, nor only seta the trash-m to zero zt those points, but also ,ets the tint derivative to zero at those points. If the intent of the i,,creaed order of polynomial is to hold down the sidelobe Ierslr, then doubling up on the zeros is a wasteful twtic. The additionnl zerca might better be placed between the existing zeroa (near the Id pelts of the sidelobes) to hold down the sidelobea rather than at locations for which the tnnaform i, already equal to zero. In fact we will observe in subsequent winders that very few good windows exhibit reprted roota. hckin.g up for a moment. it is interesting to examine the trim& window in terms of putul+un conrer~ncc of Fourier win. Fejcz observed that the pamnl sums of Fourier series were poor numericd approximations I8 I. F0lUier coefficients were ury to gwzrate however, and he questioned if some simple maditiution of coefficients might lead to a new set with mom desirable conrerynce properdes The oscillation of the partial sum, and the contraction of those os.ziUations its the order of the partiaJ sum increased. suggested that an .veng of the partiaJ surm would be a smoother function. Fii. IS prernts an expulsion of twc. partial sums near P diswntmuity. Notice the ..of the two expansions is smoother thm either. Contintxiru in this line of reasoning. an averwe expansion FN(B) mi#d be defined by F~(e)=~IF~.,(e)+F~-*(e)+~+F,(e)l
(24)

IVmm) w(n)=cos ;n , n= -i,., -I,o.I;.~ [1

(I = 1 .O (cosine lobe) N

(I = I .a (tie

lobe) n-0.1.2;..,N1

(I = 2.0

(cosine squared, raised cosine. Harming)

n=---,
a = 2.0 w(n) = sin [ =0.5 , ,:.

N
2

,-l,O,l:~~.~

(2ial

(sine squared. raised cosine, Ha&g)

1, I cw ,$?]] .

n=O,I,2;..A-

I (27b)

where FM(O) is the M<erm partial sum tif the wies~ Tti is easily in TsbL II. which lists the nonzero ccefficicnb of the fuxt four parti., sums rad their avenge summation. We see that the Fejer converaencc factors applied to the Fourier series coefficienta is, in fact. a triangle window. The averaging of partial lums is known 8?1 the method of Ccsaro summability

Thw is actuy a f&y of windows dependent upon the parameter (I, with (I nonaUy being an ir,teger. Attractmnr of this family include the ease with which the terms can be generated, and the easily identified properties of the transform

The windows are showm for (L integer values of I through 4 m Figs 16 through 19. Notice as 01 becomes larger. the wndows become rmoother and the transfom, reflects thilr increared smo~otbnmcss in decreased sidelobc level and faster falloff of :he sidelabcs. but with an increased width of the main lobe. Of particular interest in this family, is the Ham window (after the Austria,, meteorologist, Julius Von Ham,) [ 7 I. Not only is this window continuous, but so is its first derivatjve. Since the discontiuity of this window resides in the second derivative, the transfam, falls off at I/w3 or at 18 dB per octave. Let us closely emmine the transform of this window. We will gti some inteRsting insight and learn of a clever appbcation of the window under the DFT.

I . ,. .. 61 ..._

(b) (b) o*-magDi*de Of tr.rnfmln.

ne ymp,ed Haming the sequences indicated w(n) = 0.5 + 0.3 cos [

window in 2n 1 N J )

can be written

as the sum Of

_.N ..,, -l,O,l;..,

2

;-

1. (Z&a) in
t2W

Exh

sequence hu the easily recognized

[D ( e-y )

DFT indicated
+fl ( e+; )I

iv(e)=osD(e)+0.25

where

D(e)=exp(+it)

f . sin Te

[1

We recognize the DirichIet kernel at the or&in as the transform of the constant 0.5 samples and the pair of translated kernels as the trtnsfonn of the single cycle of cosine samples. Note that the translated kernels are located on the fuxt zeros of the center kernel, and arc half the size of the center kernel. Also the sideobcs of the translated kernel are about half the sizz and are of opposite phase of the sidelobes of the Central kernel. The summation of the three kernels sidelobes bcti in phse opposition. tends to cancel the sidclobc structure. m in Fig. 20 which depicts ClnceIIing summation is demonstrated the summation of the Dirichlet kernels (without the phaseshift term). The pattial cancelling of the sidelobe structure su%gssts 1 cm,str,,ctisc technique to defme new windows. The mat we&horn of these are the Hamming and the BLzckman windows which are presented in the next two sections. For the special case of the DFT. the HamUg window is sampled at multipks of 2*/N, which of course am the locations of the tcrca of the central Dirichict kernel. Thus only three nonzero sampka are taken in the SampIhlg pPxa.The positions of these samples UC ,t -2x/N, 0, and +2x/N. The due of the samples obtained from (28b) (including the phase factor exp (-j(N/2)8) to wautt for the N/2 shift) M - $, ++, - ;, rcrpectireIy. Note the minus s&us. l-hue results from the shift in the origin for the window. Without the shift. the pha term h missing and the urfticknts are all positive f , +, ;. These am infor DFT proardng. but they find their way into much of the Htentruo and practice. Rather than apply the window aa a product in the time domain, we always have the option to apply it IS a conwl:i tion in the frequency domain. The attraction of the Harming window for this application is twofold; tint. the window spectra is nonzero at only thee data pain, and second. the sample values M binary fnctioaa, xbich can be implemented as tight shifts. Thus the Haming-windowed spectral points obtained from the rectangle-windowed spectral points are obtained as indicated in the following equation as two real adds and two binary shifts (to multiply by f):
f(k)/Hdw = f [f(k) f (F(k 1) +.=(k+l)ll /R-. (29)

reai adds and 2N binary shifts on the spectral Jar;l~ One other mildly important consideration, if the window !s io be applied to the time data. is that the samples of the u,r,~.Io~~ must be stored somewhere, which normally means additional memory or hardware. It so happens that the samples of the cosine for tix Hatming window are already stored in the machine as the trig-table for the FFT; thus the Hamung window requires no additional storage.
or as 2N D. Hamming Window [7/ The Hamming window can be thought ofas a modified Hatming window. (Note the potential source of confuion m the similarities of the two name,.) Referring back to Fig.. I7 and 20, we note the inexact cancellation of the sidelobes from the summation of the three kernels. We can construct d wiw dew by adjusting the relative size of the kernels as indicated in the following to achieve a more desirable form of cancellation w(n)=u+(I -a)cos

[ 1
277 N D [ ( e-2 N) +D ( e+z N)] (3Oa)

w(e)=ao(e)+o.s(l-a)

Perfect cancellation of the occm when (I = 25146 (a = 0.54 (an approtimation to 0 A 2.612n/Nl and a marked realized. For this value of ming .window and is identified
0.54 + 0.46 cos

fmt sidelobe (at 0 = 2.5 [?n/Nl) 0.543 478 261). If a is selected a 25146). the new zero occurs af improvement in sidclobe level 1s (1. the window is called the Hamby
n

[ 1
2 N

0.54

0.46

CM

[ 1
;n ,

n=O,l.2;...N-

I.

(30b)

Thus P Harming window applied to a real transform of length N can be wrformed as N real multiplier on the tune sequence

The coefficients of the Hamming window are nearly the set wbxh a&eve mmmm sidelobe levels. If a is selected to be 0,53856 the sidelobe level is -43 dB and the resultant window LS P rpccial case of the Blackman-Harris windows presented in Sccaon V-E. The Hamming window is shown in Fig. 2 I. Vouce the deep attenuation at the missing sidelobe position. Note also that the small discontinuity at the boundary of the wmdow has resulted m a I/w (6.0 dB per octave) rate of falloff. The better sidelobe canceUation does result in a much lower initial sidelobe level of 42 dB. Table I lists the paramI.

eters of this window. Also note the loss of binary weighting; hence the need to perform multiplication to apply the weighting factors of the spectral convolution.
E. Blackman Window [7/

The Hamming and Harming windows ale examples of wirdows constmcred as the summation of shifted Dlrichlet kernels. This data window is defined for the finite Fourier transform in 131a) and ior the DFT 1n(3lb):equation(3lc)is the resultant spectral window for the DFT given as a summation of the Dtiichiet kerneis D(8) defined by W(9) I,, (2,c,;

of this form with a0 and 0, being oonzero. We see that their spectral windows are summations of three-shifted kernels. We can construct windows with any K nonzero coefficients and achieve a (2KI) summation of kernels. We recognize. however. that one way to achieve windows with a narrow main lobe is to restrict K to a small integer. Blackman examined this window for K = 3 and found the values of the nonzero coefficients which place zeros at 0 = 3.5 (?n/N) and at 0 = 45 (2nlN). the position of the third and the fourth sidelobes. respectively, of the central Dirichlet kernel. These exact values and their two place approximations are
7936 = ---a0 I8608 0.426 590 71 = 0,42

N
(3l.a)

n=O.I:....V~
m=o

1 (3lb)

9240 = --& 18608

0.496

560

62 = 0.50

~ 1430 a2 = 18608

A 0.076

848

67 = 0~08

W(O)=

x w2 ,-I) -+(+)+D(O+;m)]~ m=o ,31Cl

The window which uses these two place approximations is known as the Blackman window. When we describe this window with the exact coefficients we will refer to it as the exact Blackman window. The Blackman window is defined for the finite transform in the following equation and the window is shown in Fig. 22:
Wini=O~42+0.5Ocnr[~n] +0.08+2,]. N

Sublect to constraint ,
i

N/2 z m = 1.0~
m=o

We ian see that the Harming

and the Hamming

windows

are

=---..

The exact Blackman window is sbom in Fig. 23. The sidelobe level is 5 I dR down for the exact Blackman window anal is 58 dB down for the Blackman window. As an observation. note that the coefficients of the Blackman window sum to zero (0.42 -0.50 +0.08) at the boundaries while the exact coeiticicntn do not. Thus the Blackman window is continuous with a continuous fr% derivative at the boundary and falls off like I/w or I8 dB per octave. The exact terms (like the Hamming wmdow) have a discontinuny at the boundary and falls off like l/w or 6 dB per octave. Table I Lists the parameten of these two windows. Note that for this class of w,,down. the ao coefficxnt is the coherent gain of rhc wmdow Using a gradient search technique 191, we have found the windows which for 3. and 4.nonzero terms achieve a mintmum

sidelobe level. We have also constructed families of 3~ an,, 1~ Lerm window in which we trade ~nain-lobe width Ior SIIIC/L)II~~ level. WC caU this family the Blackman-Harris wmduu U,, have found that the minimum )-term window can achlrvr II sidelobe level of -67 dB and that the minimum d-term urns dow can acbwe a sidelobe level of -92 dB. These wmdowi are defined for the DFT by

n=O.I.::--..v-

1~ ,3?l

Thr l~tcd coeiiicients correspond to the mmm~um 3.term wmdow uhlch is presented in Fig. 24, another 3-rerm wmdou

3.Term (47 dB) 00 01 a* a3 0.42323 0.49755 0.07922

3-T(41 dB) 0.44959 0.49364 0.05677

4-TCrm (-92 **, 0.35875 0.48829 0.14128 0.01168

GTW (-74 dB, 0.40217 0.49703 0.09392 0.00183

windows (pmmetcrized on a) which were the starting condltions for the gradient minimizatioo which lads to the Bhckn~~-Hurin windows. The optimization starting with tbes mefftients baa virtually no effect on the main-lobe chancterMica but does drive down the aidelobes appmtintely 5 dB.

(to establish another data point in Fii 12). the minimum 4term window (to also eatablisb a data point in Fii. 12). md another &term window which is pxsented in Fig. 25. The particular d-term window shown is one which performs refl in a detection example described in Sxtion VI (see F& 69). The p-eters of these windows ue lined i,, Table I. Note in particular where the Blackman and the Blackman-Hanis windows reside in Fig. 12. They are surprisingly good windows for the small number of terms in their trigonometric series. Note, if we were lo extend the Line corme&ng the BlackmaeHarris family it would intereat the Hamming window which. in Section V-D . we noted is neviy the minimum sidelobe level 2-tern Blackman-Harris window. We also mentioo that agood approximation to the BlackmanHarris 3- and 4-term windows ca,, be obtained as xa,ed sampler of the Kaise-Eeszzl windows tmns(one (see Section V-H). We have used this approximation to construct b-term windows for adjustable bandwidth convolutional fdten as reponed in [ 101. This approximation is defmed as

Numennu bmst+ton have co-ted windows as products, as sums, a.3 satiom, or a.3 convolutions of simple functions and of other ximple tidoas. xx window have heen constructed for certain desirable fcshns, no, the law /.I which is the attnctioo of simple functions for genera& tiir. window terms. In gmmnf, the constructed windows tend no, to be 8c.d windows, and ocosionnlly arc very bad windows. We hare llrudy eramincd some ximplc window coostmctions. The Fejer @arUett) window. for tianoe, is the convolution of two rectangle windows; the Hamming window is the sum of a rcct+nde and a Harming window; and the cos(X, window is the product of two t&am& windows We will now enmine other mnstructed windows that have appeared m the Iitemtwe.~ We wiU present them so they me available for comptiso. later we will examine windows constructed ,n accort, with some criteria of optima,ity.(sce Sectrons VG. H. 1. an.: I). Each window is identified only for the foute Fourier trans. form A simple shift of N/2 paints and right end-pomt deletion mll supply the DFf version. The slgruficant figures of ~~rformancc for these windows are also found in Table I~
1) Rirsz (Bochner. Punen]
n lb+)= 1.0. Ii NJ2

Window
=

fll/~~

The

Raeu

XVII-

dew. identified

as
05)

ho :!!!? c The 4 coefficients

o, *o = 0.40243,

.,=2%-

c3
,,I

m=:1,2,()).

(341

when a = 3.0 are I, and a, = 0.00122. Notice how close thus terms are to the selected 4.term Blackman-Harris (- 74 dB) window. The window dcfmcd hy Ome coefficients is shown in Fig. 26. Like the prototype from which it came (the Kaiser-Bessel with a = 3.0). thus Wmdow exbihits sidelobes just shy of -70 dB from the ream lobe. On the rcale show,,. the two are indistinguisbahlr The parameters of this window are also Listed in Table I and the wmdow is entered in Fig. 12 as the 4sample Kaiserksc. If was these 3. and 4Smple Kaiser-Bessel prototype
= 0.49804. = 0.0983

for this approximation

ii thr cinlplest ~ontinuou polynominl window II exhlhits di~:ontmuoos fnt derivative at the boundaries; hrnl::: !:, tramfo fas Off bke I/w. The window is rhowvn in F&; 27. The fnt ddclobe is -22 dB from the main lohe~ This window is similar to the cosine lobe (26) IS can he demonstrated by e xammmg its Taylor series expulsion. 21 Riemmn Window 1121: The Riemann window, defined by r .
N w(n)= oGGi (36)

is the central -

lobe of the SING kernel.

This wmdow

is con-

tim,ous, with o di,cootiouous fnt derivative at the boundary. It is similar to the Ricsz md cosine lobe windows. The Ricmmn window U shown in Figs 20. 31 de in Valid-Powsin (Jackson. Punen) Window [II]. The de la VallC-PouSan window is P piecewise cubic curve obtained by self-convolving two triangles of half extent or four rectangles of one-fourth extent. It is defined as

(37:

The window is continuous up to itr third derivative so that its sidelpbes fall off like I/w. The window is shown I I:IF 20~ Notice the trade off of main-lobe width for sidelobe level. Compare this wth the rectangle and the triangle. It is a nonnaativ~ window by virtue of its self-convolution const~~tioa~. 41 Tukcy Window [13l: The Tukey window, often called the cosine-tapered window. is best imagined as a wsine lobe of width ia/Z)N convolved with a rectangle window of width I 1 .O nil)A Of course the rrsultdrrl transfurm ASthe product of the two corresponding transforms. The window rspiesznts an attempt to smoothly ret the data to zero at the bouaddtizs while not significantly reducin& the processing gam of thr wmdowed transform The window evolves from the rectangle to the Harming windou as the parameter o varies from zero lo umty. The family of windows exhibits a confusing anay 31

tidelobe Innsfons.

levels The I .o.

arising from the product window is dcti.ed by

of the

two

component

N 06116UZ

u(n)

= i 0.5

, (38) fbhe tindow is shown in Figri. 0.25. 0.50. and 0.75. respectively. 51 Bohrnon Window 1141: 3&32 The for Ekhman values of window (I equal is to o&

tained by (261). thus tmm.fohn be described cycle of a term added Thu the nutty resides I/w. The Fii 33:

convolution of two half-duration cosine lobes its transform is the square of the cosine lobzs (see Fii. 16). in the time domain tbc wndow I 11, as a product of a trim@ window with a single Mine with the same period and. the, a comective to %et the first derivative to zero at the boundary. second derivative is cmtinuous, and the discontiin the third derivative. 7%~ transform fatIs off like window is defmcd in the folloti~ and is showo in

the

6)

Poisson

Window

1121:

The

Poison

window

is P two-

sided exponential

defmed

by -a( In I O<llC~. (40)

win)=exp

N/2 )

observed in Table I as a large equivalent rroise bandwIdth and as a large worst case processing 109s. 71 Homing-Poimm Window: The Hanning-Poisson witdew U constructed as the product of the Hanrung and the Ponson window. The family is defmed by ~,,=~~s~.~+~-[=~~~~p(.~~), OC,,<fy (41) Ths window is similar to the Poisson window. sidelobe falloff is determined by the discontinuity The rate of in the first

This is actually P family of windows panmetctied on the variable a. Since it exhibits a discontinuity at the houndties. the tran?lform can fall off no faster than I/w. The window is shown in F&s 34-36 for values of d equal to 2.0. 3~0, and 4.0. rewectivcly. Notice as the discontinuity at the boundtier becomes smdlcr. fhe sidelobe strcture merges into rh: asymptote. Also no,e the very wide main lobe: this will be

derivative at the origin and is l/w. Notice as a increaser, forcing more of the exponential into the Harming window. the zeros of the nidelobc structure disappear and the lobes mrrge into the asymptote. This window is shown in Figs, 37-39 for values of (I equal to 0.5, 1.0, and 2~0. respectively. Awn note the very large main-lobe width 8: Cauchg (Abel, Poisson) Window (I5]: Tt~he Cauchy wtidow is a family parametefized on Q and defined by
w(n) = .O+ I n 1 I GJ O<,l<!~ (42)

Cauchy window is a two-sided exponent,,, (see Poisson wm, dews), which when presented on a log-mumtude scale II essentially an isoreles triangle. This C~LIYI the wincl:~ * I,, ezhibif a very wide main lobe and to have I large EEX!W.

Windows ax smooth positive functions wlh tall thin 0.c concer.trated) Fourier transforms. From the generaluc.. uncertainty principle, we know WC cannot dmultaneouslii concentrate both a signal and its Fourier transform If ou measure of concentration is the mean-square time duration :: and the mean-square bandwidth W. we know all fcncuon satisfy the inequality of

The wmdow is shown in Figs 40-42 ior valws of Q ~quzl to 3.0, 4 0. and 5~0. respectively. Note the transform of the

nv>;

143

with equality being achiclcd only for the Gau5ian ~~1161. by minimum timeThus the Gauszian pulse. characterized bandwidth product, is a rcuonablc candidate for a window. pulse as a window we havr to tnuwhen WC se the GaU ute or discard the tails. By rest&t& the puke to be finite Icwth, the window no longer is minimum time-bandwidth. If the trvnution point is beyond the thrwAgmr point, the crro would be small. and the window should be a good *PProumation to minimum time-hndwidth. The Catin window is dcf,ed by dn) = exp

n = -TClr 11
N/2

(Mb) Thu window is paameterized 0x1 a. the reciprocal of the standard deviation. a measure of the width of its Fourier tranrfomr. lnclease d a will decrease with the width of the -doe, and reduce the severity of the discontinuity at the boundaries. ThrS will result ti an increased width transform

(44)

mam lobe and decreased sidelohe levcls~ The wmdow is presented ir. F&s 43, 44. and 42 for values of a equli !o I,?, 3.0. and 3.5, respectively. Nate the rapid drop-off rate of sldelcbe lrvel in the exchange of ddelobc level for mam-lobe width. The fiurer cf merit for this wlndcw arc lured in Table I.

apenure to achieve a narrow main-lobe beam partwn wb& simultaneously restricting sidelobe rc~ponx. (The antenna drsignrr calls his weighting procedure Ihrrdinx.) The closedform solution to the minimum main-lobe width for a given sidelobe level is the Dolph-Chebyshcv window (shading) The continuous solution to the problem exhibits impulses at the bvundtics which restricts continnuous realizations to aFprOX~atlOns (the Taylor approximrfion). The discrete or sampled window is not so redtxictsd. and the solution can be unpiemented e,.actly. The relatvx~ T,C.U! = cos (n9) describes a mapping between the nth-order Cbebyshev (algebraic) polynomial and the nthwdcr trxganome:tic polynomiai. Tix Dolph4Yhebyxhe>

window is defied with this mapping in the following equation, i,, terms of u,,ifom,ly spaced runpla of the windows Fourier tnnsform.

and

* - tat- [X/&TO 2 In Ix+~Fxcil, 1 X], IX1 c 1.0 1x1> 1.0.

cos-(x)=

O<IkI<Nwhere

(45)

0=cash ; cash(IO) 1

To obtain the ccmcapooding window time samples w(n), we nmply perform a DFT on the samples W(k) and then scale for unity peak amplitude. The parameter (I reprevnts the log of the ratio of mak,in-lok level to sidelobe level. Thus a value of o equal to 3.0 rcprcscnto sidelobes 3.0 decades down from the main lobe, or sidelobes 60.0 dB below the main lobe. The (- I) al,ernates the dgn of successlre nansfoml samples to reflect the shifted origin in the time domair. The wmdow is

.,

presented in Figs. 46-49 for values of Q equal t 2.5, 3.0, 3.5, and 4.0, respectively. Note the uniformity of the sidelobe structure; almost sinusoidal! It is this uniform oscillation which is responsible for the impulses in the window.
J. Kaiser-Bessel Window /I81

The p-eta 110 is half of the time-bandwidth transform is approximately that of

product.

The

Let us examine for a mment the optimalily criteria of the last two sections. In Section V-G we sought the function with minimum time-bandwidth product. We know this to be the Gsus%iul. In Section V-H we sought the function with restricted time duration. which minimized the main-lobe width for P given sidelobe level. We now consider a simllar problem. For a restricted energy, determine the function of restricted time duration T which mnximhcs the energy in the band of frequentica W. Slepian, PoUak, and Landau ( I9 1, 1201 have determined this function as a family parametetied Ova the time-bandwidth product. the prolate-spheroidal wave functions of order zem. Kaiser hu discovered a rimplc apProximatian to these functions in terms of the zero-order modified Besel function of the fti kind. Tbe Kaise-Bessel window is defined by Jo w(n)=where

This window h presented in Fii. SO-53 for nls of acqd to 2.0. 2.5. 3.0, and 3.5, -tidy. Note the trade off between sidelobe level and main&be width.
1. Borcilon-Temes Window 1211

We now examine the lut criterion of 0ptimaLity for a window. We have already described the Slcpian, Poti, and Landau criterioo. Subject to the wnstnints of tixcd enand fued duration. determine the function which maxim&a the energy in the band of ffequcntics W. A related criterion, subject to the constraints of fued arca md fried duration. U the tllerp? (nr to determine the function which lllsmds the we&b&d energy) outside the band of frequeocies W ~#ii: is a reasonable criterion since we rxa@ze that the tnnsfon the enit gathm fmm of a good window should minimix frequencies removed from its cater frcqwy. Till now. we the onlccahave bee responding t thi, .g,al by m.ximiz@ tration of the transform at its main Lobe. A closed-form solution of the unwei&tcd minimum-encrgr criterion has not been found. A rolutia dcfmed as an cxpmsion of prolate~pheroidal ware functions dots exist sod ii is of the form shown in

Here the X,, is the e&envplue corresponding to the associated prolate-spheroidal wave function I $tLln(x. y) I. and the na is the selected half time-bandwidth product. The summation converges quite rapidly, and is often approximated by the irst term or by the tint two terms. The fust term happens to be the solution al the Slepian. Poll&, and Landau problem. which we have already examined as the Kaiser-Bessel window. A closed-form solution of a weighted minimum-energy criterion. presented in the following equation has been found by Barcilon and Tema:
Minimtie

L.ike the DolphXbcbysbw window, the Fourier truuform is more easily defmed, and the window timejlmpla an obtamed by an inverse DFT and an appropriate scale factor. Tbc transform samples are defmed by

A cos Iv(k)1 + B y? W)=(-IP

[C+ABI [p]

sin [y(L)]
+ LO]

1
(49)

IHWl&dw.

(481 between criteria. the Dolph-

where A =sinh(O=m B=cosh(C)= IO

Thiscritetion isone which is a compromise Chebyshev and the Kaiser-Bessel window

C=cosh-(IOU) @= cash y(k)= N cos

1 [1
;C

(gee alsO (45).) this window is presented in Figs. 54-56 fat vduer of a equal to 3.0, 3.5, and 4.0. respectively. The ain lobe stmctre is practically indistinguishable fro the Kaiser&s] m&+&e. me fuures of merit listed on Table 1 suggest eat f,,r the same sidelobe level, this window does indeed tide between the Kaiser-Bessel and the Dolph-Cbeby*ev h&w, it is interesting to examine Fig. 12 and note where w ,+jndow is located with respect to the Kaiser-Bessel edow; rt&ing similarity in PerfOtmanCe!

we now describe a simple experiment which dIaatic~Y demo,,strstes the influence a window exerts on the detection of a we& spectrll line in the presence of a strong nearby line. ,f two ~pccval ,i,,es reside in Dm bins, the rectangle window avows each to be identified with no interaction. ,To demonstrate this, consider the signal composed of two frequencieS ,O fi,N and 16 f,/N (corresponding to the tenth and the dteenth DFT bins) and of amplitudes 1.0 and 0.01 (40.0 dB scpustion), respectively. The power Epectru Of this Sigllal obtained by a DFT is shown in Fig. 57 as a lincar interpolation between the DFT wtput points. We now modify the signal slightly so that the larger signal resides midway between two DFT bins; in particulnr, at 10.5 f,/h'. The smaller signal still resides in the sixteenth bin. The power spectrum of this signal is show in Fig. 58. We note that the sidelobe structure of the larger signal has completely -amped the main lobe of the smaller sigznl In fact, we know (see Fig. 13) that the sidelobe amplitude of the rectangle window af 5.5 bins from the center is only 25 dB down from the peak. Thus the second signal (5.5 bins away) could not be detected because it was ore than 26 dB down. and hence, hidden by the sidelobe. (The 26 dB corns from the 25d8 sidelobe level minus the 3.9dB processing Ions of the window plus 3.0 dB for a high confidence detection.) We also note the obvious asymmetry around the main lobe centered at 10.5 bms. Thrs is due to the coherent addition of the sidelobe struucturcs of the par of kernels located at the plus and minus 10.5 bin positions. We are obsetig the self-leakage between the positive and the negative frequencies. Fig. 59 is the power spectrum of the signal pair, oditied so that the largwamplitude signal resides at the 10.25.bin position. Note the change Ln asymmetry of the main-lobe and thr rcdncti~!n in ills ridclobe level. We still can not observe the second ,.,givd 1ocatc.l a! bin position 16.0. We now apply different windows to the twxtone signal to demonstrate the difference in second-tone detectability. For SOme of the windows, the pwrer resolution OCCUIS when the large signal is at 10.0 bins rather than at 10.5 bins. We will always Present the window with the large signal at the location corresponding to warst-case resolution. The first windo* we apply is the triangle window (see Fig. 60) The side;obzs have fak,, by a factor of two over the rec*=WJc wmdowr lobes (e.g.. the -35dB ieve, has fallen to -70 dBI. The ridrlabes of the larger signal have fallen to aPPro*lnUtelY -45 dB a( the second signal so that it LI barel)

dc!tectable. If there were any noise in the signal. the second fCme would probably not have been detected. The next windows we apply are the axa family. For thmecosine lobe, a = I .O, shown in Fig. 6 I we observe a phase ca!ncellation in the sidelobe of the large signal located at the In nall signal position. T&is annnot be considered a detect W e also see the spectral leakage of the main lobe oet tue fr equency axis. Signals below this leakage level would not be d<:tectcd. With a = 2.0 we hzve the Harming wmdow, which is

Fit?. 63. ca

f.,K)

widow.

mily. For :-e a ph== zated at the a det-Xtl,be over the U&d not be ich 1s iow

presented in Fig. 62. We detect the second signal and observe 1 3.O-dB null between the two lobes. This is still a marginal detection. For the cos(x) window presented in Fig. 63, we detect the second signal and obsme a 9.O-dB null between the lobes. We al.w see the improved sidelobe response. Finally for tbe co%(x) window presented in Fig. 64, we detect the second signal and ob~rve a 7.O-dB null between the lobes. Here WC witnes the reduced return for the trade between sidelobe level and main-lobe width. In obtaining further reduction in sidelobe level we have caused tbc increased mati~8Obe width to encroach upon the second SignaL t We next apply the Hamming window and present the result h.. FS 65 Here we observe the second signal some 35 dB am. approximately 3.0 dB over the sidelobe response of e be siamI. Here. too, we observe the phase cancellation (I the leakage between the positive and the negative freeCY components. Signals more than 50 dB down would & t be detected in the ;;resenze of the larger signal. Thp BLaclwaa window is applied next and we szn the revolts 2 Fig. 66. Tbe presence of the smaller amplitude kernel is now WV apparent. There ir a 17-dB null between the two dCnak IX: artifact at the base of the lqe-signal kernel is

I!
i!

1;
.

the sidelobe stmcture of tJut kernel. Note the rapid rate of falloff of the sidelobe Leakage has confined the artifacts to a small portion of the spectral line. We next apply the exact Blackman coefficients and witness the results :;n Fig. 67. Again the second Ugnal is well defined with a 24-ziB null between the two kernels. The sidelobe strctre of the larger kernel now extends over the entire spectral range. This leafrage is not terribly severe as it is nearly

P-t&70. .amptc---rila.

a.

73. de I# vstt.5-Polli0

window

to the phuc can-tion of a sidelobe in the law &I kernel. The result of a Riemam window is presented in FU. 72. Here. too, we bare no detuztion of the second s&ml. WC do have a muU aull due to piuse rmwlfation at the second sw ad. We also have a Loge ddelobe rapome. The ,,ext window. the de la Vail&Poussin or the rlfconvolved triangle. is shown io Fig 73. Tbe second n.qnnll u a.dy found lad the power spectrum exhibits a 16.O-dB null. An artifact of the window (itz lower sidelobe) shorn up. hoaewr, at the fiith DFT bin as a signal approximately 53.0 dB down See Fy 29. The rcdt of applyins the Tukcy family of window ix praeoted in Fm. 74-76. In Fig. 74 (the 25percent taper) WC lice the hck of ,ecmd+wal detection due to the b&b sidelobe smxturc of the dominant rectangle window. In Fi&. 75 (the 50perccnt taper) WC observe a lack of second-ngml detection. with the second s%?mal actually filling in one of the oulk of the furt s&ml.% kernel. In Fig. 76 (the 76-percent taper) we witness a marginal detection in the still high side-.. .

The B&mm construction window is applied and presented in Fia. 77. The second signal has been detected and :hv null between thr two 1&es is approximately 6.0 dB. Thk is not bad. but we can still do better. Note rbere the Bohman wi,~ dew resides m Fig. 12. The result of applying the Poisson-window famdy 1s presm:ed i:. Fig.. 7.?-50~ Ti-x second :Unal is nor detccrsd f@r any of the selected parameter values due to the higkidrlobe

levels of the larger signal. We anticipated tha poor performance in Table I by the large differmce betrem the 3.0 dB and the ENBW. The result of applying the Hanning-Poisson family of windows is presented in Figs. 81-83. Here. too. the second signal is either not detected in the presence of UIC hi&Jidelobe structure or the detection is bewildered by the utifncts. Tb.e Cauchy-family windows have been apphcd and tl~,i results are presented in Figr. 84-86. Here too we havr a Ia< k of satisfactory detection of the second s&nll and tkrc poor sidelobe mporse. Tbis was predicted by the lqc difference between the 3.0 dB and the cquivaknt noise bandwidths as listed in Table I. We now apply the Gausian family of windows and present the results in F@. 87-89. The second si.mnl L detected in aU three fwen. we note a we further depress the sidelobr strulic:,ue to enhance secondsi~a, detection, the null deepens to approximately 16.0 dB and then becomes poorer as the main-lobe width increases and starts to overlap the lobe of the smdler signal. The DolphXhebyzhtv family of windows is presented in Figs. 90-94. We observe strong detection of the second signal

lobe being sampled off of the peak and being refcrenwd as MD dB. Fii. 90 and 91 dcmoostnte tic sensitivity of the sidelobe coherent addition to main&& position. In Fig. 90 the larger sigh is at bin L0.5; in Fig 91 it is at bin 10.0.

in phase cancellation near the base of the the 7MBsidelobe window. exhibits an 1 MB null between the two main lobes but the sidelobes have added constructively (along with the scalloping loss) to the -62.O-dB level. In Fig. 94, we see the 8WB udelobe -0oU exhibited sidelobes below the 7048 level and still managed to bold the null between the two lobes to approximatley 18.0 dB. The KG.ser-Bessel family is presented in Figs 95-98. Here.
Large s&nab Fig 93.

Note the difference

too. we have strong second-signal detection. Again. we see the effect of trading increased main-lobe width for decreased sldelobe level. The null between the two lobes reaches a maximum of 22.0 dB as the sidelobe stru~~ture falls and then becomes poorer with further sidelobe level improvement Note that this window can maintab, a 20.OdB null between the two Wnal lobes and sti hold the leakage to more than 70 dB down over thr entire spectrum. Figs. 99-101 present the performance of the BarcilonTews window. Note the strong detection of the second s,gnal.

There are slight sidelobe artifacts. The window can maintain a 20,OdB null between the two signal lobes. The performance of this window is slightly shy of that of the Kaiser-Bessel wmdow. but the two are remarkably similar. VII. CoNCLUsloNs

We have examined some classic windows and some windows which satisfy some criteria of optimality. In particular. we have desctihed their effects on the problem of general har-

0.*.m _

1 i\ j; ! :

manic analysis of tones in broadband noise and of tones in the presence of other tones. We have observed that when the DFT Is used as a hvmonic energy detector, the worst case procesing 1os1 due to the windows sppearr to be lower bounded by 3.0 dB and (far good windows) upper bounded near 3.75 dB. This suggesta that the choice of particular windows has very little effect on worst case performance in DFl energy detection. We have concluded that a good performance indicator for the window 15 the difference between the equivalent noise bandwidth and the 3.048 bandwidth normalized by the 3.o-dB bandwidth. The windows which perform weU (as indicated in Fig. 12) exhibit vallres for this ratio between 4.0 and 5.5 percent. The range of this ratio for the windows listed in Table I is 3.2 to 22.9 percent. For multiple-tone detection via the DFT, the window employed does have a considerable effect. Maximum dynamic range of multitone detection requires the transform of the window to exhibit a highly concentrated central lobe with very-low sidelobe structure. We have demonstrated that many classic windam satisfy this miterion with varymg

degrees of suxess and some not at all. We have demonsr; .,, the optimal windows (Kaiser-Bessel. Dolph-Chebyshcv, aid Barcilon-Temes) and the Blackman-Hanis windows perform best in detection of nearby tones of dgnificantly differmr amplitudes. Also for the same dynamic range, the three optimal windows and the Blackman-Harris window are roughly equivalent with the Kaiser-Bessel and the Blackman-Hams. demonstrating minor perfornmnce advantages over the others We note that while the Dolphxhebyshev window appears to be the best window by virtue of its relative position in Fig. 12. the coherent addition of its constant-level sidelobes detracts from its perfomwtce in multi tone detection. Also the PIIIClobe st,,,ctwe of fhe Dolph-Chebyshev window rxhilvt\ extreme sensitivity to coefficient errors This would affect its performace in machines operating with fixed-point arithmetic. This suggests that the Kaiser-Bessel or the BlackmanHarris window be declared the top performer. My preference is the Kaiser-Bessel window. Among other reasons. the curiticienfs are easy to generate and the tradeall of sidelob? level as a function of time-bandtidth product is fairly ample, For many applications, the author would recommend thr 4~ sample Blackman-Harris (or the 4sample Kaiser-Brssrl) window. These have the distinctiw~ czf being defini+ by :+ :.: easily generatcd coefficients and of being able to ix apl., li ar a spectral convolution after the DFT. We have called attention to a petistent error in the application of windows when performing convolution in the frequency domain. i.e., the omission of the alternating signs on the window sample spectrum to account for the shifted time Ogln. We have also identified and clarified a source of confusion concerning the e~ennesz of windows under the DFT. Finally. we comment that all of the conclusions presented about window performance in spectral analysis are also applicable to shading for array processing of spatial sampled data, including FFT beamforming.

HARRIS:

SE

OF WINDOWS

FOR

HARMONlC

ANAYSIS

*P f(l) and Yw)= Then

F,(w) = x n-w(nT)jC(nT) exp (+jwnT)

F(w)

exp

(-jwr)

dwl2n

+Nl, x .-V,*

w(nT)

exp (+jwnT).

becomes
IS F,(w)= 1

F(x) exp (-jxnT)dxjZn exp (+jwnT)

l fW

n=-

F(x) --

x =a

w(nT)

exp

I+j(w

x)nTl

drj2n

*a = F(x) / _m
.m =I-

*TV,*
x =-IV,* w(nT) exp Ifj (w - x)nTl dx/2n

F(x) W(w

x)dx/2n

or F,(W) = F(W) * W(W).